Daniel Ju
What I build, how I work, and the direction I am pursuing.
The short version
I build and operate ML research pipelines, agent infrastructure, and the self-hosted platform behind this site. I completed Computer Science and Political Science at George Washington in May 2026. I begin Georgia Tech’s OMSCS machine learning specialization in Fall 2026.
Two languages
Computer Science and Political Science shape how I work. In 2026 I completed a CKA interpretability study across eight model architectures and a 78-page econometric analysis of USPTO patent geography under the CHIPS Act. I use technical methods in policy research and institutional context in engineering.
The patent direction
At GWU’s Technology Commercialization Office, I translate licensed university inventions and patent portfolios into public communications and support invention-disclosure intake. My Islands of Innovation research works with patent data, CPC taxonomy, and inventor geography at scale. I am preparing for the patent bar with a focus on ML, AI, and semiconductor IP.
ML/AI, honestly measured
My research emphasizes evaluation controls. The Convergence Mirage measures representation similarity against untrained-network floors. Swap or Flop checks whether strong RL benchmark results correspond to strategic behavior. DELAMAIN applies the same discipline to production agent infrastructure through replayable streams, deterministic tool gates, and an enforced privacy boundary.
The ecosystem
I run an always-on Linux server and a GPU workstation with Cloudflare tunnels and Access, Tailscale, Syncthing, nginx, and systemd. This site is served from that stack. The Systems page documents how the public services, private network, and deployment path fit together.